2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
25 #include "midgard_ops.h"
26 #include "midgard_quirks.h"
27 #include "util/u_memory.h"
28 #include "util/u_math.h"
29 #include "util/half_float.h"
31 /* Scheduling for Midgard is complicated, to say the least. ALU instructions
32 * must be grouped into VLIW bundles according to following model:
35 * [VADD] [SMUL] [VLUT]
37 * A given instruction can execute on some subset of the units (or a few can
38 * execute on all). Instructions can be either vector or scalar; only scalar
39 * instructions can execute on SADD/SMUL units. Units on a given line execute
40 * in parallel. Subsequent lines execute separately and can pass results
41 * directly via pipeline registers r24/r25, bypassing the register file.
43 * A bundle can optionally have 128-bits of embedded constants, shared across
44 * all of the instructions within a bundle.
46 * Instructions consuming conditionals (branches and conditional selects)
47 * require their condition to be written into the conditional register (r31)
48 * within the same bundle they are consumed.
50 * Fragment writeout requires its argument to be written in full within the
51 * same bundle as the branch, with no hanging dependencies.
53 * Load/store instructions are also in bundles of simply two instructions, and
54 * texture instructions have no bundling.
56 * -------------------------------------------------------------------------
60 /* We create the dependency graph with per-byte granularity */
65 add_dependency(struct util_dynarray
*table
, unsigned index
, uint16_t mask
, midgard_instruction
**instructions
, unsigned child
)
67 for (unsigned i
= 0; i
< BYTE_COUNT
; ++i
) {
68 if (!(mask
& (1 << i
)))
71 struct util_dynarray
*parents
= &table
[(BYTE_COUNT
* index
) + i
];
73 util_dynarray_foreach(parents
, unsigned, parent
) {
74 BITSET_WORD
*dependents
= instructions
[*parent
]->dependents
;
76 /* Already have the dependency */
77 if (BITSET_TEST(dependents
, child
))
80 BITSET_SET(dependents
, child
);
81 instructions
[child
]->nr_dependencies
++;
87 mark_access(struct util_dynarray
*table
, unsigned index
, uint16_t mask
, unsigned parent
)
89 for (unsigned i
= 0; i
< BYTE_COUNT
; ++i
) {
90 if (!(mask
& (1 << i
)))
93 util_dynarray_append(&table
[(BYTE_COUNT
* index
) + i
], unsigned, parent
);
98 mir_create_dependency_graph(midgard_instruction
**instructions
, unsigned count
, unsigned node_count
)
100 size_t sz
= node_count
* BYTE_COUNT
;
102 struct util_dynarray
*last_read
= calloc(sizeof(struct util_dynarray
), sz
);
103 struct util_dynarray
*last_write
= calloc(sizeof(struct util_dynarray
), sz
);
105 for (unsigned i
= 0; i
< sz
; ++i
) {
106 util_dynarray_init(&last_read
[i
], NULL
);
107 util_dynarray_init(&last_write
[i
], NULL
);
110 /* Initialize dependency graph */
111 for (unsigned i
= 0; i
< count
; ++i
) {
112 instructions
[i
]->dependents
=
113 calloc(BITSET_WORDS(count
), sizeof(BITSET_WORD
));
115 instructions
[i
]->nr_dependencies
= 0;
118 /* Populate dependency graph */
119 for (signed i
= count
- 1; i
>= 0; --i
) {
120 if (instructions
[i
]->compact_branch
)
123 unsigned dest
= instructions
[i
]->dest
;
124 unsigned mask
= mir_bytemask(instructions
[i
]);
126 mir_foreach_src((*instructions
), s
) {
127 unsigned src
= instructions
[i
]->src
[s
];
129 if (src
< node_count
) {
130 unsigned readmask
= mir_bytemask_of_read_components(instructions
[i
], src
);
131 add_dependency(last_write
, src
, readmask
, instructions
, i
);
135 if (dest
< node_count
) {
136 add_dependency(last_read
, dest
, mask
, instructions
, i
);
137 add_dependency(last_write
, dest
, mask
, instructions
, i
);
138 mark_access(last_write
, dest
, mask
, i
);
141 mir_foreach_src((*instructions
), s
) {
142 unsigned src
= instructions
[i
]->src
[s
];
144 if (src
< node_count
) {
145 unsigned readmask
= mir_bytemask_of_read_components(instructions
[i
], src
);
146 mark_access(last_read
, src
, readmask
, i
);
151 /* If there is a branch, all instructions depend on it, as interblock
152 * execution must be purely in-order */
154 if (instructions
[count
- 1]->compact_branch
) {
155 BITSET_WORD
*dependents
= instructions
[count
- 1]->dependents
;
157 for (signed i
= count
- 2; i
>= 0; --i
) {
158 if (BITSET_TEST(dependents
, i
))
161 BITSET_SET(dependents
, i
);
162 instructions
[i
]->nr_dependencies
++;
166 /* Free the intermediate structures */
167 for (unsigned i
= 0; i
< sz
; ++i
) {
168 util_dynarray_fini(&last_read
[i
]);
169 util_dynarray_fini(&last_write
[i
]);
176 /* Does the mask cover more than a scalar? */
179 is_single_component_mask(unsigned mask
)
183 for (int c
= 0; c
< 8; ++c
) {
188 return components
== 1;
191 /* Helpers for scheudling */
194 mir_is_scalar(midgard_instruction
*ains
)
196 /* Do we try to use it as a vector op? */
197 if (!is_single_component_mask(ains
->mask
))
200 /* Otherwise, check mode hazards */
201 bool could_scalar
= true;
202 unsigned szd
= nir_alu_type_get_type_size(ains
->dest_type
);
203 unsigned sz0
= nir_alu_type_get_type_size(ains
->src_types
[0]);
204 unsigned sz1
= nir_alu_type_get_type_size(ains
->src_types
[1]);
206 /* Only 16/32-bit can run on a scalar unit */
207 could_scalar
&= (szd
== 16) || (szd
== 32);
209 if (ains
->src
[0] != ~0)
210 could_scalar
&= (sz0
== 16) || (sz0
== 32);
212 if (ains
->src
[1] != ~0)
213 could_scalar
&= (sz1
== 16) || (sz1
== 32);
218 /* How many bytes does this ALU instruction add to the bundle? */
221 bytes_for_instruction(midgard_instruction
*ains
)
223 if (ains
->unit
& UNITS_ANY_VECTOR
)
224 return sizeof(midgard_reg_info
) + sizeof(midgard_vector_alu
);
225 else if (ains
->unit
== ALU_ENAB_BRANCH
)
226 return sizeof(midgard_branch_extended
);
227 else if (ains
->compact_branch
)
228 return sizeof(ains
->br_compact
);
230 return sizeof(midgard_reg_info
) + sizeof(midgard_scalar_alu
);
233 /* We would like to flatten the linked list of midgard_instructions in a bundle
234 * to an array of pointers on the heap for easy indexing */
236 static midgard_instruction
**
237 flatten_mir(midgard_block
*block
, unsigned *len
)
239 *len
= list_length(&block
->base
.instructions
);
244 midgard_instruction
**instructions
=
245 calloc(sizeof(midgard_instruction
*), *len
);
249 mir_foreach_instr_in_block(block
, ins
)
250 instructions
[i
++] = ins
;
255 /* The worklist is the set of instructions that can be scheduled now; that is,
256 * the set of instructions with no remaining dependencies */
259 mir_initialize_worklist(BITSET_WORD
*worklist
, midgard_instruction
**instructions
, unsigned count
)
261 for (unsigned i
= 0; i
< count
; ++i
) {
262 if (instructions
[i
]->nr_dependencies
== 0)
263 BITSET_SET(worklist
, i
);
267 /* Update the worklist after an instruction terminates. Remove its edges from
268 * the graph and if that causes any node to have no dependencies, add it to the
273 BITSET_WORD
*worklist
, unsigned count
,
274 midgard_instruction
**instructions
, midgard_instruction
*done
)
276 /* Sanity check: if no instruction terminated, there is nothing to do.
277 * If the instruction that terminated had dependencies, that makes no
278 * sense and means we messed up the worklist. Finally, as the purpose
279 * of this routine is to update dependents, we abort early if there are
280 * no dependents defined. */
285 assert(done
->nr_dependencies
== 0);
287 if (!done
->dependents
)
290 /* We have an instruction with dependents. Iterate each dependent to
291 * remove one dependency (`done`), adding dependents to the worklist
295 BITSET_FOREACH_SET(i
, done
->dependents
, count
) {
296 assert(instructions
[i
]->nr_dependencies
);
298 if (!(--instructions
[i
]->nr_dependencies
))
299 BITSET_SET(worklist
, i
);
302 free(done
->dependents
);
305 /* While scheduling, we need to choose instructions satisfying certain
306 * criteria. As we schedule backwards, we choose the *last* instruction in the
307 * worklist to simulate in-order scheduling. Chosen instructions must satisfy a
308 * given predicate. */
310 struct midgard_predicate
{
311 /* TAG or ~0 for dont-care */
314 /* True if we want to pop off the chosen instruction */
317 /* For ALU, choose only this unit */
320 /* State for bundle constants. constants is the actual constants
321 * for the bundle. constant_count is the number of bytes (up to
322 * 16) currently in use for constants. When picking in destructive
323 * mode, the constants array will be updated, and the instruction
324 * will be adjusted to index into the constants array */
326 midgard_constants
*constants
;
327 unsigned constant_mask
;
330 /* Exclude this destination (if not ~0) */
333 /* Don't schedule instructions consuming conditionals (since we already
334 * scheduled one). Excludes conditional branches and csel */
337 /* Require (or reject) a minimal mask and (if nonzero) given
338 * destination. Used for writeout optimizations */
344 /* For VADD/VLUT whether to only/never schedule imov/fmov instructions
345 * This allows non-move instructions to get priority on each unit */
348 /* For load/store: how many pipeline registers are in use? The two
349 * scheduled instructions cannot use more than the 256-bits of pipeline
350 * space available or RA will fail (as it would run out of pipeline
351 * registers and fail to spill without breaking the schedule) */
353 unsigned pipeline_count
;
357 mir_adjust_constant(midgard_instruction
*ins
, unsigned src
,
358 unsigned *bundle_constant_mask
,
359 unsigned *comp_mapping
,
360 uint8_t *bundle_constants
,
363 unsigned type_size
= nir_alu_type_get_type_size(ins
->src_types
[src
]) / 8;
364 unsigned type_shift
= util_logbase2(type_size
);
365 unsigned max_comp
= mir_components_for_type(ins
->src_types
[src
]);
366 unsigned comp_mask
= mir_from_bytemask(mir_round_bytemask_up(
367 mir_bytemask_of_read_components_index(ins
, src
),
370 unsigned type_mask
= (1 << type_size
) - 1;
372 /* Upper only makes sense for 16-bit */
373 if (type_size
!= 16 && upper
)
376 /* For 16-bit, we need to stay on either upper or lower halves to avoid
377 * disrupting the swizzle */
378 unsigned start
= upper
? 8 : 0;
379 unsigned length
= (type_size
== 2) ? 8 : 16;
381 for (unsigned comp
= 0; comp
< max_comp
; comp
++) {
382 if (!(comp_mask
& (1 << comp
)))
385 uint8_t *constantp
= ins
->constants
.u8
+ (type_size
* comp
);
386 unsigned best_reuse_bytes
= 0;
387 signed best_place
= -1;
390 for (i
= start
; i
< (start
+ length
); i
+= type_size
) {
391 unsigned reuse_bytes
= 0;
393 for (j
= 0; j
< type_size
; j
++) {
394 if (!(*bundle_constant_mask
& (1 << (i
+ j
))))
396 if (constantp
[j
] != bundle_constants
[i
+ j
])
398 if ((i
+ j
) > (start
+ length
))
404 /* Select the place where existing bytes can be
405 * reused so we leave empty slots to others
407 if (j
== type_size
&&
408 (reuse_bytes
> best_reuse_bytes
|| best_place
< 0)) {
409 best_reuse_bytes
= reuse_bytes
;
415 /* This component couldn't fit in the remaining constant slot,
416 * no need check the remaining components, bail out now
421 memcpy(&bundle_constants
[i
], constantp
, type_size
);
422 *bundle_constant_mask
|= type_mask
<< best_place
;
423 comp_mapping
[comp
] = best_place
>> type_shift
;
429 /* For an instruction that can fit, adjust it to fit and update the constants
430 * array, in destructive mode. Returns whether the fitting was successful. */
433 mir_adjust_constants(midgard_instruction
*ins
,
434 struct midgard_predicate
*pred
,
437 /* Blend constants dominate */
438 if (ins
->has_blend_constant
) {
439 if (pred
->constant_mask
)
441 else if (destructive
) {
442 pred
->blend_constant
= true;
443 pred
->constant_mask
= 0xffff;
448 /* No constant, nothing to adjust */
449 if (!ins
->has_constants
)
452 unsigned r_constant
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
453 unsigned bundle_constant_mask
= pred
->constant_mask
;
454 unsigned comp_mapping
[2][16] = { };
455 uint8_t bundle_constants
[16];
457 memcpy(bundle_constants
, pred
->constants
, 16);
459 /* Let's try to find a place for each active component of the constant
462 for (unsigned src
= 0; src
< 2; ++src
) {
463 if (ins
->src
[src
] != SSA_FIXED_REGISTER(REGISTER_CONSTANT
))
466 /* First, try lower half (or whole for !16) */
467 if (mir_adjust_constant(ins
, src
, &bundle_constant_mask
,
468 comp_mapping
[src
], bundle_constants
, false))
471 /* Next, try upper half */
472 if (mir_adjust_constant(ins
, src
, &bundle_constant_mask
,
473 comp_mapping
[src
], bundle_constants
, true))
480 /* If non-destructive, we're done */
484 /* Otherwise update the constant_mask and constant values */
485 pred
->constant_mask
= bundle_constant_mask
;
486 memcpy(pred
->constants
, bundle_constants
, 16);
488 /* Use comp_mapping as a swizzle */
489 mir_foreach_src(ins
, s
) {
490 if (ins
->src
[s
] == r_constant
)
491 mir_compose_swizzle(ins
->swizzle
[s
], comp_mapping
[s
], ins
->swizzle
[s
]);
497 /* Conservative estimate of the pipeline registers required for load/store */
500 mir_pipeline_count(midgard_instruction
*ins
)
502 unsigned bytecount
= 0;
504 mir_foreach_src(ins
, i
) {
505 /* Skip empty source */
506 if (ins
->src
[i
] == ~0) continue;
508 unsigned bytemask
= mir_bytemask_of_read_components_index(ins
, i
);
510 unsigned max
= util_logbase2(bytemask
) + 1;
514 return DIV_ROUND_UP(bytecount
, 16);
517 /* Matches FADD x, x with modifiers compatible. Since x + x = x * 2, for
518 * any x including of the form f(y) for some swizzle/abs/neg function f */
521 mir_is_add_2(midgard_instruction
*ins
)
523 if (ins
->alu
.op
!= midgard_alu_op_fadd
)
526 if (ins
->src
[0] != ins
->src
[1])
529 if (ins
->src_types
[0] != ins
->src_types
[1])
532 for (unsigned i
= 0; i
< MIR_VEC_COMPONENTS
; ++i
) {
533 if (ins
->swizzle
[0][i
] != ins
->swizzle
[1][i
])
537 if (ins
->src_abs
[0] != ins
->src_abs
[1])
540 if (ins
->src_neg
[0] != ins
->src_neg
[1])
547 mir_adjust_unit(midgard_instruction
*ins
, unsigned unit
)
549 /* FADD x, x = FMUL x, #2 */
550 if (mir_is_add_2(ins
) && (unit
& (UNITS_MUL
| UNIT_VLUT
))) {
551 ins
->alu
.op
= midgard_alu_op_fmul
;
554 ins
->src_abs
[1] = false;
555 ins
->src_neg
[1] = false;
557 ins
->has_inline_constant
= true;
558 ins
->inline_constant
= _mesa_float_to_half(2.0);
563 mir_has_unit(midgard_instruction
*ins
, unsigned unit
)
565 if (alu_opcode_props
[ins
->alu
.op
].props
& unit
)
568 /* FADD x, x can run on any adder or any multiplier */
569 if (mir_is_add_2(ins
))
575 static midgard_instruction
*
576 mir_choose_instruction(
577 midgard_instruction
**instructions
,
578 BITSET_WORD
*worklist
, unsigned count
,
579 struct midgard_predicate
*predicate
)
581 /* Parse the predicate */
582 unsigned tag
= predicate
->tag
;
583 bool alu
= tag
== TAG_ALU_4
;
584 bool ldst
= tag
== TAG_LOAD_STORE_4
;
585 unsigned unit
= predicate
->unit
;
586 bool branch
= alu
&& (unit
== ALU_ENAB_BR_COMPACT
);
587 bool scalar
= (unit
!= ~0) && (unit
& UNITS_SCALAR
);
588 bool no_cond
= predicate
->no_cond
;
590 unsigned mask
= predicate
->mask
;
591 unsigned dest
= predicate
->dest
;
592 bool needs_dest
= mask
& 0xF;
594 /* Iterate to find the best instruction satisfying the predicate */
597 signed best_index
= -1;
598 bool best_conditional
= false;
600 /* Enforce a simple metric limiting distance to keep down register
601 * pressure. TOOD: replace with liveness tracking for much better
604 unsigned max_active
= 0;
605 unsigned max_distance
= 6;
607 BITSET_FOREACH_SET(i
, worklist
, count
) {
608 max_active
= MAX2(max_active
, i
);
611 BITSET_FOREACH_SET(i
, worklist
, count
) {
612 bool is_move
= alu
&&
613 (instructions
[i
]->alu
.op
== midgard_alu_op_imov
||
614 instructions
[i
]->alu
.op
== midgard_alu_op_fmov
);
616 if ((max_active
- i
) >= max_distance
)
619 if (tag
!= ~0 && instructions
[i
]->type
!= tag
)
622 if (predicate
->exclude
!= ~0 && instructions
[i
]->dest
== predicate
->exclude
)
625 if (alu
&& !branch
&& !(mir_has_unit(instructions
[i
], unit
)))
628 if ((unit
== UNIT_VLUT
|| unit
== UNIT_VADD
) && (predicate
->moves
!= is_move
))
631 if (branch
&& !instructions
[i
]->compact_branch
)
634 if (alu
&& scalar
&& !mir_is_scalar(instructions
[i
]))
637 if (alu
&& !mir_adjust_constants(instructions
[i
], predicate
, false))
640 if (needs_dest
&& instructions
[i
]->dest
!= dest
)
643 if (mask
&& ((~instructions
[i
]->mask
) & mask
))
646 if (instructions
[i
]->mask
& predicate
->no_mask
)
649 if (ldst
&& mir_pipeline_count(instructions
[i
]) + predicate
->pipeline_count
> 2)
652 bool conditional
= alu
&& !branch
&& OP_IS_CSEL(instructions
[i
]->alu
.op
);
653 conditional
|= (branch
&& instructions
[i
]->branch
.conditional
);
655 if (conditional
&& no_cond
)
658 /* Simulate in-order scheduling */
659 if ((signed) i
< best_index
)
663 best_conditional
= conditional
;
667 /* Did we find anything? */
672 /* If we found something, remove it from the worklist */
673 assert(best_index
< count
);
675 if (predicate
->destructive
) {
676 BITSET_CLEAR(worklist
, best_index
);
679 mir_adjust_constants(instructions
[best_index
], predicate
, true);
682 predicate
->pipeline_count
+= mir_pipeline_count(instructions
[best_index
]);
685 mir_adjust_unit(instructions
[best_index
], unit
);
687 /* Once we schedule a conditional, we can't again */
688 predicate
->no_cond
|= best_conditional
;
691 return instructions
[best_index
];
694 /* Still, we don't choose instructions in a vacuum. We need a way to choose the
695 * best bundle type (ALU, load/store, texture). Nondestructive. */
699 midgard_instruction
**instructions
,
700 BITSET_WORD
*worklist
, unsigned count
)
702 /* At the moment, our algorithm is very simple - use the bundle of the
703 * best instruction, regardless of what else could be scheduled
704 * alongside it. This is not optimal but it works okay for in-order */
706 struct midgard_predicate predicate
= {
708 .destructive
= false,
712 midgard_instruction
*chosen
= mir_choose_instruction(instructions
, worklist
, count
, &predicate
);
720 /* We want to choose an ALU instruction filling a given unit */
722 mir_choose_alu(midgard_instruction
**slot
,
723 midgard_instruction
**instructions
,
724 BITSET_WORD
*worklist
, unsigned len
,
725 struct midgard_predicate
*predicate
,
728 /* Did we already schedule to this slot? */
732 /* Try to schedule something, if not */
733 predicate
->unit
= unit
;
734 *slot
= mir_choose_instruction(instructions
, worklist
, len
, predicate
);
736 /* Store unit upon scheduling */
737 if (*slot
&& !((*slot
)->compact_branch
))
738 (*slot
)->unit
= unit
;
741 /* When we are scheduling a branch/csel, we need the consumed condition in the
742 * same block as a pipeline register. There are two options to enable this:
744 * - Move the conditional into the bundle. Preferred, but only works if the
745 * conditional is used only once and is from this block.
746 * - Copy the conditional.
748 * We search for the conditional. If it's in this block, single-use, and
749 * without embedded constants, we schedule it immediately. Otherwise, we
750 * schedule a move for it.
752 * mir_comparison_mobile is a helper to find the moveable condition.
756 mir_comparison_mobile(
757 compiler_context
*ctx
,
758 midgard_instruction
**instructions
,
759 struct midgard_predicate
*predicate
,
763 if (!mir_single_use(ctx
, cond
))
768 for (unsigned i
= 0; i
< count
; ++i
) {
769 if (instructions
[i
]->dest
!= cond
)
772 /* Must fit in an ALU bundle */
773 if (instructions
[i
]->type
!= TAG_ALU_4
)
776 /* If it would itself require a condition, that's recursive */
777 if (OP_IS_CSEL(instructions
[i
]->alu
.op
))
780 /* We'll need to rewrite to .w but that doesn't work for vector
781 * ops that don't replicate (ball/bany), so bail there */
783 if (GET_CHANNEL_COUNT(alu_opcode_props
[instructions
[i
]->alu
.op
].props
))
786 /* Ensure it will fit with constants */
788 if (!mir_adjust_constants(instructions
[i
], predicate
, false))
791 /* Ensure it is written only once */
799 /* Inject constants now that we are sure we want to */
801 mir_adjust_constants(instructions
[ret
], predicate
, true);
806 /* Using the information about the moveable conditional itself, we either pop
807 * that condition off the worklist for use now, or create a move to
808 * artificially schedule instead as a fallback */
810 static midgard_instruction
*
811 mir_schedule_comparison(
812 compiler_context
*ctx
,
813 midgard_instruction
**instructions
,
814 struct midgard_predicate
*predicate
,
815 BITSET_WORD
*worklist
, unsigned count
,
816 unsigned cond
, bool vector
, unsigned *swizzle
,
817 midgard_instruction
*user
)
819 /* TODO: swizzle when scheduling */
821 (!vector
&& (swizzle
[0] == 0)) ?
822 mir_comparison_mobile(ctx
, instructions
, predicate
, count
, cond
) : ~0;
824 /* If we can, schedule the condition immediately */
825 if ((comp_i
!= ~0) && BITSET_TEST(worklist
, comp_i
)) {
826 assert(comp_i
< count
);
827 BITSET_CLEAR(worklist
, comp_i
);
828 return instructions
[comp_i
];
831 /* Otherwise, we insert a move */
833 midgard_instruction mov
= v_mov(cond
, cond
);
834 mov
.mask
= vector
? 0xF : 0x1;
835 memcpy(mov
.swizzle
[1], swizzle
, sizeof(mov
.swizzle
[1]));
837 return mir_insert_instruction_before(ctx
, user
, mov
);
840 /* Most generally, we need instructions writing to r31 in the appropriate
843 static midgard_instruction
*
844 mir_schedule_condition(compiler_context
*ctx
,
845 struct midgard_predicate
*predicate
,
846 BITSET_WORD
*worklist
, unsigned count
,
847 midgard_instruction
**instructions
,
848 midgard_instruction
*last
)
850 /* For a branch, the condition is the only argument; for csel, third */
851 bool branch
= last
->compact_branch
;
852 unsigned condition_index
= branch
? 0 : 2;
854 /* csel_v is vector; otherwise, conditions are scalar */
855 bool vector
= !branch
&& OP_IS_CSEL_V(last
->alu
.op
);
857 /* Grab the conditional instruction */
859 midgard_instruction
*cond
= mir_schedule_comparison(
860 ctx
, instructions
, predicate
, worklist
, count
, last
->src
[condition_index
],
861 vector
, last
->swizzle
[2], last
);
863 /* We have exclusive reign over this (possibly move) conditional
864 * instruction. We can rewrite into a pipeline conditional register */
866 predicate
->exclude
= cond
->dest
;
867 cond
->dest
= SSA_FIXED_REGISTER(31);
870 cond
->mask
= (1 << COMPONENT_W
);
872 mir_foreach_src(cond
, s
) {
873 if (cond
->src
[s
] == ~0)
876 for (unsigned q
= 0; q
< 4; ++q
)
877 cond
->swizzle
[s
][q
+ COMPONENT_W
] = cond
->swizzle
[s
][q
];
881 /* Schedule the unit: csel is always in the latter pipeline, so a csel
882 * condition must be in the former pipeline stage (vmul/sadd),
883 * depending on scalar/vector of the instruction itself. A branch must
884 * be written from the latter pipeline stage and a branch condition is
885 * always scalar, so it is always in smul (exception: ball/bany, which
889 cond
->unit
= UNIT_SMUL
;
891 cond
->unit
= vector
? UNIT_VMUL
: UNIT_SADD
;
896 /* Schedules a single bundle of the given type */
898 static midgard_bundle
899 mir_schedule_texture(
900 midgard_instruction
**instructions
,
901 BITSET_WORD
*worklist
, unsigned len
,
904 struct midgard_predicate predicate
= {
905 .tag
= TAG_TEXTURE_4
,
910 midgard_instruction
*ins
=
911 mir_choose_instruction(instructions
, worklist
, len
, &predicate
);
913 mir_update_worklist(worklist
, len
, instructions
, ins
);
915 struct midgard_bundle out
= {
916 .tag
= ins
->texture
.op
== TEXTURE_OP_BARRIER
?
917 TAG_TEXTURE_4_BARRIER
: is_vertex
?
918 TAG_TEXTURE_4_VTX
: TAG_TEXTURE_4
,
919 .instruction_count
= 1,
920 .instructions
= { ins
}
926 static midgard_bundle
928 midgard_instruction
**instructions
,
929 BITSET_WORD
*worklist
, unsigned len
)
931 struct midgard_predicate predicate
= {
932 .tag
= TAG_LOAD_STORE_4
,
937 /* Try to pick two load/store ops. Second not gauranteed to exist */
939 midgard_instruction
*ins
=
940 mir_choose_instruction(instructions
, worklist
, len
, &predicate
);
942 midgard_instruction
*pair
=
943 mir_choose_instruction(instructions
, worklist
, len
, &predicate
);
945 struct midgard_bundle out
= {
946 .tag
= TAG_LOAD_STORE_4
,
947 .instruction_count
= pair
? 2 : 1,
948 .instructions
= { ins
, pair
}
951 /* We have to update the worklist atomically, since the two
952 * instructions run concurrently (TODO: verify it's not pipelined) */
954 mir_update_worklist(worklist
, len
, instructions
, ins
);
955 mir_update_worklist(worklist
, len
, instructions
, pair
);
961 mir_schedule_zs_write(
962 compiler_context
*ctx
,
963 struct midgard_predicate
*predicate
,
964 midgard_instruction
**instructions
,
965 BITSET_WORD
*worklist
, unsigned len
,
966 midgard_instruction
*branch
,
967 midgard_instruction
**smul
,
968 midgard_instruction
**vadd
,
969 midgard_instruction
**vlut
,
972 bool success
= false;
973 unsigned idx
= stencil
? 3 : 2;
974 unsigned src
= (branch
->src
[0] == ~0) ? SSA_FIXED_REGISTER(1) : branch
->src
[idx
];
976 predicate
->dest
= src
;
977 predicate
->mask
= 0x1;
979 midgard_instruction
**units
[] = { smul
, vadd
, vlut
};
980 unsigned unit_names
[] = { UNIT_SMUL
, UNIT_VADD
, UNIT_VLUT
};
982 for (unsigned i
= 0; i
< 3; ++i
) {
986 predicate
->unit
= unit_names
[i
];
987 midgard_instruction
*ins
=
988 mir_choose_instruction(instructions
, worklist
, len
, predicate
);
991 ins
->unit
= unit_names
[i
];
998 predicate
->dest
= predicate
->mask
= 0;
1003 midgard_instruction
*mov
= ralloc(ctx
, midgard_instruction
);
1004 *mov
= v_mov(src
, make_compiler_temp(ctx
));
1007 branch
->src
[idx
] = mov
->dest
;
1010 unsigned swizzle
= (branch
->src
[0] == ~0) ? COMPONENT_Y
: COMPONENT_X
;
1012 for (unsigned c
= 0; c
< 16; ++c
)
1013 mov
->swizzle
[1][c
] = swizzle
;
1016 for (unsigned i
= 0; i
< 3; ++i
) {
1017 if (!(*(units
[i
]))) {
1019 mov
->unit
= unit_names
[i
];
1024 unreachable("Could not schedule Z/S move to any unit");
1027 static midgard_bundle
1029 compiler_context
*ctx
,
1030 midgard_instruction
**instructions
,
1031 BITSET_WORD
*worklist
, unsigned len
)
1033 struct midgard_bundle bundle
= {};
1035 unsigned bytes_emitted
= sizeof(bundle
.control
);
1037 struct midgard_predicate predicate
= {
1039 .destructive
= true,
1041 .constants
= &bundle
.constants
1044 midgard_instruction
*vmul
= NULL
;
1045 midgard_instruction
*vadd
= NULL
;
1046 midgard_instruction
*vlut
= NULL
;
1047 midgard_instruction
*smul
= NULL
;
1048 midgard_instruction
*sadd
= NULL
;
1049 midgard_instruction
*branch
= NULL
;
1051 mir_choose_alu(&branch
, instructions
, worklist
, len
, &predicate
, ALU_ENAB_BR_COMPACT
);
1052 mir_update_worklist(worklist
, len
, instructions
, branch
);
1053 unsigned writeout
= branch
? branch
->writeout
: 0;
1055 if (branch
&& branch
->branch
.conditional
) {
1056 midgard_instruction
*cond
= mir_schedule_condition(ctx
, &predicate
, worklist
, len
, instructions
, branch
);
1058 if (cond
->unit
== UNIT_VADD
)
1060 else if (cond
->unit
== UNIT_SMUL
)
1063 unreachable("Bad condition");
1066 /* If we have a render target reference, schedule a move for it. Since
1067 * this will be in sadd, we boost this to prevent scheduling csel into
1070 if (writeout
&& (branch
->constants
.u32
[0] || ctx
->is_blend
)) {
1071 sadd
= ralloc(ctx
, midgard_instruction
);
1072 *sadd
= v_mov(~0, make_compiler_temp(ctx
));
1073 sadd
->unit
= UNIT_SADD
;
1075 sadd
->has_inline_constant
= true;
1076 sadd
->inline_constant
= branch
->constants
.u32
[0];
1077 branch
->src
[1] = sadd
->dest
;
1078 branch
->src_types
[1] = sadd
->dest_type
;
1080 /* Mask off any conditionals. Could be optimized to just scalar
1081 * conditionals TODO */
1082 predicate
.no_cond
= true;
1087 bundle
.last_writeout
= branch
->last_writeout
;
1090 /* When MRT is in use, writeout loops require r1.w to be filled (with a
1091 * return address? by symmetry with Bifrost, etc), at least for blend
1092 * shaders to work properly. When MRT is not in use (including on SFBD
1093 * GPUs), this is not needed. Blend shaders themselves don't know if
1094 * they are paired with MRT or not so they always need this, at least
1097 if (writeout
&& (ctx
->is_blend
|| ctx
->writeout_branch
[1])) {
1098 vadd
= ralloc(ctx
, midgard_instruction
);
1099 *vadd
= v_mov(~0, make_compiler_temp(ctx
));
1101 if (!ctx
->is_blend
) {
1102 vadd
->alu
.op
= midgard_alu_op_iadd
;
1103 vadd
->src
[0] = SSA_FIXED_REGISTER(31);
1104 vadd
->src_types
[0] = nir_type_uint32
;
1106 for (unsigned c
= 0; c
< 16; ++c
)
1107 vadd
->swizzle
[0][c
] = COMPONENT_X
;
1109 vadd
->has_inline_constant
= true;
1110 vadd
->inline_constant
= 0;
1112 vadd
->src
[1] = SSA_FIXED_REGISTER(1);
1113 vadd
->src_types
[0] = nir_type_uint32
;
1115 for (unsigned c
= 0; c
< 16; ++c
)
1116 vadd
->swizzle
[1][c
] = COMPONENT_W
;
1119 vadd
->unit
= UNIT_VADD
;
1121 branch
->dest
= vadd
->dest
;
1122 branch
->dest_type
= vadd
->dest_type
;
1125 if (writeout
& PAN_WRITEOUT_Z
)
1126 mir_schedule_zs_write(ctx
, &predicate
, instructions
, worklist
, len
, branch
, &smul
, &vadd
, &vlut
, false);
1128 if (writeout
& PAN_WRITEOUT_S
)
1129 mir_schedule_zs_write(ctx
, &predicate
, instructions
, worklist
, len
, branch
, &smul
, &vadd
, &vlut
, true);
1131 mir_choose_alu(&smul
, instructions
, worklist
, len
, &predicate
, UNIT_SMUL
);
1133 for (unsigned moves
= 0; moves
< 2; ++moves
) {
1134 predicate
.moves
= moves
;
1135 predicate
.no_mask
= writeout
? (1 << 3) : 0;
1136 mir_choose_alu(&vlut
, instructions
, worklist
, len
, &predicate
, UNIT_VLUT
);
1137 predicate
.no_mask
= 0;
1138 mir_choose_alu(&vadd
, instructions
, worklist
, len
, &predicate
, UNIT_VADD
);
1141 mir_update_worklist(worklist
, len
, instructions
, vlut
);
1142 mir_update_worklist(worklist
, len
, instructions
, vadd
);
1143 mir_update_worklist(worklist
, len
, instructions
, smul
);
1145 bool vadd_csel
= vadd
&& OP_IS_CSEL(vadd
->alu
.op
);
1146 bool smul_csel
= smul
&& OP_IS_CSEL(smul
->alu
.op
);
1148 if (vadd_csel
|| smul_csel
) {
1149 midgard_instruction
*ins
= vadd_csel
? vadd
: smul
;
1150 midgard_instruction
*cond
= mir_schedule_condition(ctx
, &predicate
, worklist
, len
, instructions
, ins
);
1152 if (cond
->unit
== UNIT_VMUL
)
1154 else if (cond
->unit
== UNIT_SADD
)
1157 unreachable("Bad condition");
1160 /* Stage 2, let's schedule sadd before vmul for writeout */
1161 mir_choose_alu(&sadd
, instructions
, worklist
, len
, &predicate
, UNIT_SADD
);
1163 /* Check if writeout reads its own register */
1166 midgard_instruction
*stages
[] = { sadd
, vadd
, smul
, vlut
};
1167 unsigned src
= (branch
->src
[0] == ~0) ? SSA_FIXED_REGISTER(0) : branch
->src
[0];
1168 unsigned writeout_mask
= 0x0;
1169 bool bad_writeout
= false;
1171 for (unsigned i
= 0; i
< ARRAY_SIZE(stages
); ++i
) {
1175 if (stages
[i
]->dest
!= src
)
1178 writeout_mask
|= stages
[i
]->mask
;
1179 bad_writeout
|= mir_has_arg(stages
[i
], branch
->src
[0]);
1182 /* It's possible we'll be able to schedule something into vmul
1183 * to fill r0. Let's peak into the future, trying to schedule
1184 * vmul specially that way. */
1186 unsigned full_mask
= 0xF;
1188 if (!bad_writeout
&& writeout_mask
!= full_mask
) {
1189 predicate
.unit
= UNIT_VMUL
;
1190 predicate
.dest
= src
;
1191 predicate
.mask
= writeout_mask
^ full_mask
;
1193 struct midgard_instruction
*peaked
=
1194 mir_choose_instruction(instructions
, worklist
, len
, &predicate
);
1198 vmul
->unit
= UNIT_VMUL
;
1199 writeout_mask
|= predicate
.mask
;
1200 assert(writeout_mask
== full_mask
);
1204 predicate
.dest
= predicate
.mask
= 0;
1207 /* Finally, add a move if necessary */
1208 if (bad_writeout
|| writeout_mask
!= full_mask
) {
1209 unsigned temp
= (branch
->src
[0] == ~0) ? SSA_FIXED_REGISTER(0) : make_compiler_temp(ctx
);
1211 vmul
= ralloc(ctx
, midgard_instruction
);
1212 *vmul
= v_mov(src
, temp
);
1213 vmul
->unit
= UNIT_VMUL
;
1214 vmul
->mask
= full_mask
^ writeout_mask
;
1216 /* Rewrite to use our temp */
1218 for (unsigned i
= 0; i
< ARRAY_SIZE(stages
); ++i
) {
1220 mir_rewrite_index_dst_single(stages
[i
], src
, temp
);
1223 mir_rewrite_index_src_single(branch
, src
, temp
);
1227 mir_choose_alu(&vmul
, instructions
, worklist
, len
, &predicate
, UNIT_VMUL
);
1229 mir_update_worklist(worklist
, len
, instructions
, vmul
);
1230 mir_update_worklist(worklist
, len
, instructions
, sadd
);
1232 bundle
.has_blend_constant
= predicate
.blend_constant
;
1233 bundle
.has_embedded_constants
= predicate
.constant_mask
!= 0;
1235 unsigned padding
= 0;
1237 /* Now that we have finished scheduling, build up the bundle */
1238 midgard_instruction
*stages
[] = { vmul
, sadd
, vadd
, smul
, vlut
, branch
};
1240 for (unsigned i
= 0; i
< ARRAY_SIZE(stages
); ++i
) {
1242 bundle
.control
|= stages
[i
]->unit
;
1243 bytes_emitted
+= bytes_for_instruction(stages
[i
]);
1244 bundle
.instructions
[bundle
.instruction_count
++] = stages
[i
];
1246 /* If we branch, we can't spill to TLS since the store
1247 * instruction will never get executed. We could try to
1248 * break the bundle but this is probably easier for
1252 stages
[i
]->no_spill
|= (1 << REG_CLASS_WORK
);
1256 /* Pad ALU op to nearest word */
1258 if (bytes_emitted
& 15) {
1259 padding
= 16 - (bytes_emitted
& 15);
1260 bytes_emitted
+= padding
;
1263 /* Constants must always be quadwords */
1264 if (bundle
.has_embedded_constants
)
1265 bytes_emitted
+= 16;
1267 /* Size ALU instruction for tag */
1268 bundle
.tag
= (TAG_ALU_4
) + (bytes_emitted
/ 16) - 1;
1270 /* MRT capable GPUs use a special writeout procedure */
1271 if (writeout
&& !(ctx
->quirks
& MIDGARD_NO_UPPER_ALU
))
1274 bundle
.padding
= padding
;
1275 bundle
.control
|= bundle
.tag
;
1280 /* Schedule a single block by iterating its instruction to create bundles.
1281 * While we go, tally about the bundle sizes to compute the block size. */
1285 schedule_block(compiler_context
*ctx
, midgard_block
*block
)
1287 /* Copy list to dynamic array */
1289 midgard_instruction
**instructions
= flatten_mir(block
, &len
);
1294 /* Calculate dependencies and initial worklist */
1295 unsigned node_count
= ctx
->temp_count
+ 1;
1296 mir_create_dependency_graph(instructions
, len
, node_count
);
1298 /* Allocate the worklist */
1299 size_t sz
= BITSET_WORDS(len
) * sizeof(BITSET_WORD
);
1300 BITSET_WORD
*worklist
= calloc(sz
, 1);
1301 mir_initialize_worklist(worklist
, instructions
, len
);
1303 struct util_dynarray bundles
;
1304 util_dynarray_init(&bundles
, NULL
);
1306 block
->quadword_count
= 0;
1307 unsigned blend_offset
= 0;
1310 unsigned tag
= mir_choose_bundle(instructions
, worklist
, len
);
1311 midgard_bundle bundle
;
1313 if (tag
== TAG_TEXTURE_4
)
1314 bundle
= mir_schedule_texture(instructions
, worklist
, len
, ctx
->stage
!= MESA_SHADER_FRAGMENT
);
1315 else if (tag
== TAG_LOAD_STORE_4
)
1316 bundle
= mir_schedule_ldst(instructions
, worklist
, len
);
1317 else if (tag
== TAG_ALU_4
)
1318 bundle
= mir_schedule_alu(ctx
, instructions
, worklist
, len
);
1322 util_dynarray_append(&bundles
, midgard_bundle
, bundle
);
1324 if (bundle
.has_blend_constant
)
1325 blend_offset
= block
->quadword_count
;
1327 block
->quadword_count
+= midgard_tag_props
[bundle
.tag
].size
;
1330 /* We emitted bundles backwards; copy into the block in reverse-order */
1332 util_dynarray_init(&block
->bundles
, block
);
1333 util_dynarray_foreach_reverse(&bundles
, midgard_bundle
, bundle
) {
1334 util_dynarray_append(&block
->bundles
, midgard_bundle
, *bundle
);
1336 util_dynarray_fini(&bundles
);
1338 /* Blend constant was backwards as well. blend_offset if set is
1339 * strictly positive, as an offset of zero would imply constants before
1340 * any instructions which is invalid in Midgard. TODO: blend constants
1341 * are broken if you spill since then quadword_count becomes invalid
1345 ctx
->blend_constant_offset
= ((ctx
->quadword_count
+ block
->quadword_count
) - blend_offset
- 1) * 0x10;
1347 block
->scheduled
= true;
1348 ctx
->quadword_count
+= block
->quadword_count
;
1350 /* Reorder instructions to match bundled. First remove existing
1351 * instructions and then recreate the list */
1353 mir_foreach_instr_in_block_safe(block
, ins
) {
1354 list_del(&ins
->link
);
1357 mir_foreach_instr_in_block_scheduled_rev(block
, ins
) {
1358 list_add(&ins
->link
, &block
->base
.instructions
);
1361 free(instructions
); /* Allocated by flatten_mir() */
1366 midgard_schedule_program(compiler_context
*ctx
)
1368 midgard_promote_uniforms(ctx
);
1370 /* Must be lowered right before scheduling */
1371 mir_squeeze_index(ctx
);
1372 mir_lower_special_reads(ctx
);
1373 mir_squeeze_index(ctx
);
1375 /* Lowering can introduce some dead moves */
1377 mir_foreach_block(ctx
, _block
) {
1378 midgard_block
*block
= (midgard_block
*) _block
;
1379 midgard_opt_dead_move_eliminate(ctx
, block
);
1380 schedule_block(ctx
, block
);